本論文我們提出了新的多模型語者確認架構,主要的研究方向是結合高斯混合模型(Gaussian Mixture Model, GMM)、支撐向量機模型(Support Vector Machine, SVM)與模糊模型(Fuzzy Model)以對傳統的單一模型方式之語者確認系統做進一步的辨識性能改良。 在高斯混合模型與支撐向量機模型之整合的多模型之語者辨認架構中,我們提出了平行式整合及序列式整合兩種機制,其分別為Voting-GMMSVM及GMM-dependent SVM 等兩種語者辨識方法。所提出之兩個方法經過三種語音資料庫的實驗測試得以驗證以有效性。與傳統高斯混合模型及支撐向量機分類器相較之下,Voting-GMMSVM之76.27%及GMM-dependent SVM之77.41%的辨識效能皆有大幅度的上升且具備競爭力。 在支撐向量機模型、模糊模型與高斯混合模型之整合的多模型之語者辨認架構中,我們提出了FDoMV&ID-SVM辨識方法,該方法藉由模糊控制器之依據合法語者與仿冒語者之兩類高斯混合模型之模型平均向量差異及相似度分數差異來調整SVM的邊界大小進而提昇SVM分類器的辨識準確度。實驗結果證實所提出之FDoMV&ID-SVM方法得以再改善傳統SVM方法之辨識性能。 所發展之強化SVM的FDoMV&ID-SVM方法更進一步再導入於前述所提出之Voting-GMMSVM與GMM-dependent SVM等高斯混合模型與支撐向量機模型混合方法以達到最大化的辨識效能。所導入之方式即是直接將此兩種混合方法中之SVM分類器直接替換為FDoMV&ID-SVM分類器。由實驗結果得知在強化SVM後之新的多模型之語者確認架構於強化之Voting-GMMSVM及強化之GMM-dependent SVM各有著78.51% 和81.84%之較為準確的辨識率,實驗結果亦證實了此類經強化SVM分類器後之多模型架構的有效性。
In this thesis, we present a new multi-model speaker recognition framework. The main purpose of this thesis is to combine the GMM model, the SVM model and the fuzzy model to enhance the performance of the conventional single model speaker verification scheme. In the speaker verification framework of GMM and SVM dual-model combination, we present the parallel-style and serial-style model combination methods, which are Voting-GMMSVM and GMM-dependent SVM, respectively. Both of the proposed methods can be validated to be effective from the experimental results. Compared with conventional SVM-based and GMM-based speaker verification, the recognition rats of the developed Voting-GMMSVM and GMM-dependent SVM are a little more satisfactory, which achieve the performance of 76.27% and 77.41%, respectively. In addition, we present the FDoMV&ID-SVM method to combine SVM, GMM and fuzzy models. This developed method is to use the distance of mean vectors and the difference of likelihood scores between valid speakers and invalid speakers GMM models to build the fuzzy model. The experimental results show that the proposed FDoMV&ID-SVM can improve the recognition performance of conventional SVM. Furthermore, the improved SVM classifier, the FDoMV&ID-SVM, is further integrated into the multi-model speaker verification framework. The SVM classifiers of Voting-GMMSVM and GMM-dependent SVM are directly replaced with the FDoMV&ID-SVM classifiers. The recognition performances of improved Voting-GMMSVM and GMM-dependent SVM multi-model frameworks are improved, which are 78.51% and 81.84%, respectively. The experiment results prove the validness of the developed multi-model speaker verification.